Skip to content
All articles
Guides

How to train an AI chatbot on your FAQs, PDFs and website

Overhead view of a desk with printed FAQ sheets, a product catalogue and notes being sorted into piles

When a small business “trains” an AI chatbot on its FAQs, PDFs and website, it is almost never training a model. It is giving a general-purpose language model a library to look things up in. When a customer asks a question, the software searches your content for the most relevant passages, hands those passages to the model, and the model writes an answer from them. The technical name is retrieval-augmented generation, or RAG, and AWS describes it as a way to make a model reference an authoritative knowledge base outside its training data, without retraining the model.

That changes what doing a good job looks like. The quality of your bot depends far more on what you write down, and how you write it, than on which vendor you pick. This guide covers what to gather, how to write answers an AI can use, how to test, which rules to add, and a 60-minute checklist to get a first version live.

What “training” a chatbot on your own data actually means

People use the word training for two very different things:

Approach What happens Good for Practical for a small business?
Fine-tuning The model itself is adjusted using example conversations Teaching a style or a narrow, repetitive task at scale Rarely. It needs many good examples, and facts learned this way are hard to update.
Retrieval (RAG) Your content is indexed, and relevant passages are fetched for each question Answering from facts that change: prices, hours, policies Yes. Edit a document and the next answer uses the new version.

Three consequences follow from retrieval, and they explain most bad chatbot answers:

  • If it isn’t written down, the bot doesn’t know it. A policy that only exists as “what we usually do” will either be missing or guessed.
  • If two sources disagree, the bot may use either. An old price list sitting next to the new one is a coin toss.
  • Retrieval finds passages, not whole documents. An answer buried in paragraph nine of a 40-page PDF, phrased in legal language, is harder to find than a short, direct answer.

What content to gather first

Start with what customers already ask, not with everything you have. The best source of questions is your own inbox. Scroll back through the last 50 to 100 customer emails, chat messages or call notes and write down each distinct question. You will usually find that a fairly short list of topics covers most of them.

Then collect the content that answers those questions:

  • Your existing FAQ page, even if it’s short or out of date.
  • Policies customers ask about: returns, refunds, cancellations, deposits, warranties, delivery.
  • Prices and what’s included, with conditions such as minimum orders, seasonal rates and extras.
  • Opening hours, holiday hours, locations, parking and directions.
  • Plain-language descriptions of your services or products, including who each one is for and who it isn’t for.
  • How to book, order, pay and reach a person.

Leave out anything the bot shouldn’t repeat to a stranger: internal procedures, supplier terms, staff notes, drafts, old price lists kept “for reference”, and any document containing customer personal data. If you wouldn’t publish it on your website, don’t upload it.

How to write answers an AI can use

Most businesses already have content. The problem is that it was written for people browsing a page, not for a system that grabs one passage at a time. A few habits make a large difference.

One question per answer

Write each answer so it stands alone. If the retrieval step pulls just that paragraph, it should still make sense without the heading above it or the paragraph before it. “Yes, see above” is useless to a bot.

Be specific: numbers, dates, conditions

Vague content produces vague answers, or worse, confident guesses. Compare these two versions of a returns answer:

Weak: We offer flexible returns on most items.

Useful: You can return unused items within 30 days of delivery for a full refund. Sale items and personalised products can’t be returned. The customer pays return shipping unless the item arrived damaged. Refunds go back to the original payment method within 5 to 7 business days of us receiving the parcel.

The second version answers five follow-up questions before they’re asked. It also gives the bot the exceptions, which is where most wrong answers come from.

Write down the “no” answers

Customers ask about things you don’t do. If your content never mentions them, a language model may fill the gap with something plausible. Add short, explicit answers: “We don’t offer same-day delivery.” “We don’t accept pets, except assistance dogs.” “We can’t quote custom work over chat; we give a price after a short call.”

Use the customer’s words

If customers type “cancel my booking” and your policy page says “reservation modification terms”, retrieval may still work, but you make it easier by using the words people actually use. Put common alternatives in the question itself: “Can I cancel or change my booking?”

Put dates on anything that expires

Seasonal offers, holiday hours and temporary changes should say when they apply: “From 20 December to 3 January we’re open 10:00 to 16:00.” Undated temporary content is a common source of answers that were true once.

How to format documents so they’re read correctly

Upload tools extract the text from your files before indexing it. What looks fine on screen can come out as a jumble. A few rules help:

  • Use real text, not scans. A PDF that is a photo of a printed page may contain no extractable text at all. If you can’t select the words with your cursor, the bot probably can’t read them either.
  • Use headings and short sections. One topic per section, with a heading that says what it covers.
  • Keep tables simple. A price list with one row per item and clear column names extracts well. Merged cells, multi-level headers and tables split across pages often don’t. For long price lists, a spreadsheet or CSV is usually cleaner than a designed PDF.
  • Avoid multi-column layouts. Brochure-style PDFs with text in two or three columns can be extracted in the wrong reading order.
  • Keep one current version. Delete the old file when you upload the new one.

Website pages vs uploaded files

Most tools let you add website pages as sources, upload files, and write answers directly. Each suits different content.

Source Best for Watch out for
Website pages Public content you already maintain: FAQs, service pages, policies Menus, cookie banners and footer text mixed in; pages that load content with JavaScript may be read as empty; check whether the tool re-reads pages automatically or only when you ask
Uploaded files (PDF, Word, Excel, CSV, TXT) Menus, price lists, detailed policies, product specs Files go stale quietly because nobody remembers to re-upload after a change
Written Q&A entries The short, specific answers described above Duplicating what’s on your site, then updating only one copy

A practical split: keep your website as the source of truth for anything public, and write Q&A entries for the specific answers your site doesn’t cover well. Upload files when the content is genuinely document-shaped, like a menu or a spec sheet.

Test with real questions, not your own

Owners test chatbots with questions phrased the way the owner thinks. Customers don’t write like that. Before going live, take 20 to 30 real questions from your inbox and paste them in exactly as they were written, typos included. Make sure the set includes:

  • the five questions you get most often;
  • two or three with missing context, such as “are you open tomorrow?”;
  • a couple that ask two things in one message;
  • one or two in another language, if you serve customers from abroad;
  • questions the bot should decline or hand over, such as a refund dispute.

For each answer, check three things: is it correct, did it come from the right source, and would you be comfortable if a customer took a screenshot of it? Some tools show which sources an answer used. In KuraChat, the dashboard’s test chat does this, which makes it quicker to tell whether a wrong answer came from outdated content or from no content at all.

When an answer is wrong, fix the content first. Rewriting the source so it is clear and specific solves most problems. Rules are for the cases content can’t handle.

Add rules and guardrails

Rules are short instructions that override the rest of your content. They hold business judgment that doesn’t belong in a FAQ. Useful examples:

  • “Never promise a delivery date. Give the standard delivery window and link to tracking.”
  • “Don’t give medical, legal or financial advice. Suggest booking an appointment.”
  • “Hand refund disputes and complaints to a person.”
  • “Don’t discuss competitors or their prices.”
  • “If our content doesn’t cover a question, say so and offer to connect the customer with the team.”

That last rule matters most. A bot that says “I don’t know, let me get someone” is mildly annoying. A bot that invents a policy can cost real money. In Moffatt v. Air Canada (2024), a Canadian tribunal held the airline liable after its website chatbot told a customer he could apply for a bereavement fare retroactively, which contradicted the airline’s actual policy. The tribunal rejected the argument that the chatbot was responsible for its own statements. Treat anything your bot says as something your business said.

Also make it obvious that it’s an AI. Customers shouldn’t have to guess whether they’re talking to a person, and in the EU it’s a legal requirement. Our AI transparency page explains how we approach disclosure.

Keep it up to date

A chatbot knowledge base decays the same way a website does, just less visibly. Three habits keep it accurate:

  1. Name an owner. One person is responsible for the bot’s content, even if it only takes them 20 minutes a week.
  2. Tie updates to business changes. When prices, hours, menus or policies change, updating the bot is part of the change, the same as updating the website.
  3. Review what it couldn’t answer. Once a week, read the questions the bot handled badly and write answers for the ones that repeat. This is how the knowledge base grows in the right direction.

A 60-minute setup checklist

  1. Minutes 0 to 10: Pull 30 real customer questions from your inbox, chat history or call notes into one list.
  2. Minutes 10 to 20: Add your core business details: what you do, where, opening hours, contact options, how to book or buy.
  3. Minutes 20 to 35: Write direct answers to the ten most common questions, one question per answer, with numbers, conditions and exceptions.
  4. Minutes 35 to 40: Add your key website pages (FAQ, pricing, policies) and upload any current menu, price list or spec sheet. Delete outdated versions.
  5. Minutes 40 to 45: Write three to five rules: what to hand over, what never to promise, what to do when unsure.
  6. Minutes 45 to 55: Test with the 30 real questions. Mark each answer right, wrong or missing.
  7. Minutes 55 to 60: Fix the two or three worst gaps, check that customers can reach a person, and put a weekly 20-minute review in your calendar.

You won’t cover everything in an hour, and you don’t need to. A bot that answers the top ten questions accurately and hands over the rest is more useful than one that attempts everything and gets some of it wrong.

Where KuraChat fits

KuraChat works the way this guide describes. It answers from your business details, written FAQs, website pages you add as sources and, on the Pro plan, uploaded PDF, Word, Excel, CSV and TXT files. Your own AI rules override everything else, the test chat shows which sources each answer used, and customers can ask for a person at any point. You can see how it works on the features page, or start on the free plan from the pricing page.

Frequently asked questions

Do I need to retrain the AI when my prices change?

No. With a retrieval-based chatbot you update the source content (the page, file or written answer) and the next answer uses the new version. Remove the old version so the two don't conflict.

How much content do I need before going live?

Enough to answer your ten most common questions accurately, plus a clear way to reach a person for everything else. Add more from the questions the bot can't answer in its first few weeks.

Can I just point the chatbot at my website?

You can, and it's a reasonable start if your site is detailed and current. Most sites are written for browsing, so you'll usually need to add specific answers for policies, exceptions and questions the site doesn't cover.

Is my uploaded content used to train the AI model?

It depends on the vendor and the model provider, so ask directly and read their data processing terms. In a retrieval setup your content is sent to the model as context for each answer, which is different from being used to train it.

KT
KuraChat team
HYPER TRELLIS S.R.L., Bucharest

Written and fact-checked by the team that builds KuraChat, AI customer support for small businesses. Sources are linked throughout the article.

About KuraChat